Foobot Adds AI to Make Any Building Smart

Foobot is a trailblazer in the traditionally conservative building management industry. Its business model and core market have changed dramatically in the five years it’s been operating, moving from the domestic to the commercial marketplace. But it has never owned on-premises hardware—Amazon Web Services (AWS) has always provided its technology setup. Even its previous domestic business allowed it to take advantage of AWS scale—at one point it had 150,000 air quality sensors connected at the same time.

Foobot uses AWS Machine Learning combined with its own smart sensors to optimize office temperature and air quality, while also reducing energy consumption. It’s an example of a company—and a business model—that couldn’t exist without the flex and scale of cloud computing. It relies on short bursts of extremely intense computing activity, which is all but impossible to manage cost-effectively with a traditional IT data center.

When an office or building manager comes to Foobot for help, the company builds a digital twin—a virtual model of the office or workspace. Then it creates an artificial intelligence agent that learns the best way to run the heating, cooling, and air conditioning systems for the virtual building. This allows Foobot to have a trained agent ready to run real-world systems in a matter of days, instead of needing to first spend months collecting data.

After the agent is trained and in place, running a building’s systems, a period of intense machine learning follows as the virtual agent takes control of the building’s air and heating systems and then fine-tunes settings by reacting to feedback from building sensors. Foobot’s solution can use a building’s existing sensors, but customers can also choose to install a network of Foobot devices. After the agent has been fully trained in how to optimize a building’s temperature and air quality, the computing demands are drastically reduced. The company has tested running a trained agent on a single Raspberry Pi mini-computer and found it could execute demands in less than a second.

In comparison, the learning and modeling phases require between one and three instances, each with 48 CPUs, on Amazon Elastic Compute Cloud (Amazon EC2).

AWS allows Foobot to run this complex system with just two technical staff members who are responsible for the learning system. It also means the company can scale up its systems quickly when bringing new customers on board: Foobot can easily manage the massive compute and storage workload of simulating a building, training a virtual agent, and processing sensor data, and can then rapidly scale back after this intense compute period is complete.

Foobot uses Amazon EC2 Image Builder running on Amazon EC2 servers for the distributed learning process.

After they are trained, the agents run in Docker containers on Amazon Elastic Container Service (Amazon ECS). Agent decisions are sent via a bridge to the building’s control systems. All of this is automated and managed through an API.

Antoine Galataud, head of data engineering at Foobot, says: “There are so many services that we rely on—we could not be where we are without AWS because we are a small team... But Amazon DynamoDB, which just runs with no maintenance, is probably the most important for us.”

Data analytics is vitally important to Foobot’s products and their future development. The company’s use of virtual environments to train its agents is a key differentiator from its competitors. Foobot uses Amazon Simple Storage Service (Amazon S3) to create a data lake and analyzes data directly using Amazon Athena.

It uses AWS Glue for creating catalogues, and Amazon SageMaker for analysis with Jupyter Notebooks.

Foobot uses data analytics both internally and externally and at every stage of every project. This enables better training of agents and improvements in sensor analysis, and powers internal dashboards for monitoring building performance. The company also opens data to customers, via both an open API and customer dashboards.

As Foobot’s customer numbers and associated data volumes grow, it expects its simulations to become increasingly accurate and its agents to become ever quicker to train.

Looking ahead, Foobot is assessing use of Amazon SageMaker RL, which could help to manage its learning environments.

Early pilots in Denmark and UK have shown that Foobot users see dramatic improvements in office environmental quality, as well as reductions in energy consumption. One construction company using Foobot technology was able to reduce energy consumption at its headquarters by up to 52 percent.

While facilities managers often choose to provide oversight and supervision and validate decisions during the first few days, Foobot’s AI agents can work effectively with no human input.

Foobot currently runs its building control agents on the cloud, but its next step is to move to an edge computing infrastructure. This will allow its agents to run on a relatively cheap box in the actual building and will help reduce concerns about potential disruptions to internet access. For customers, having the technology on site could also improve their perceptions about security.

Looking ahead, Foobot also hopes to make increasing use of data analytics. This will help reduce the time it takes to move new customers from initial inquiry to having active agents in place. It will also make it easier for Foobot to take agents that are pre-trained in one environment and apply those to similar environments in other locations. This process should become faster and easier as training and simulation improve and Foobot’s dataset grows. By keeping data at the center of everything it does, Foobot aims to continue helping customers reduce costs and gain ever-better control over building heating, cooling, and ventilation budgets.